Researchers

Rahaf Al Sharif, Nada Ghniem and Oumayma Al Dakkak

Published in

Damascus University Journal for Engineering Sciences, First conference on Informatics, volume 38, issue 4, October 2022.

Abstract

The multi-label emotion classification task aims to identify all possible emotions in a written text that best represent the person's mental state. It tries to understand his expressions and emotions in the text, including feelings (sadness, anger, disgust, surprise, fear, joy, ....). Recently, multi-label emotion

classification MLEC attracted the attention of researchers due to its potential applications in e-learning, healthcare, marketing, etc. As we aim to integrate this system in a speech synthesis application, our work will be on Standard Arabic sentences. We study 11 emotions, namely: (anger, joy, sadness, surprise, disgust, anticipation, love, pessimism, fear, optimism, confidence). This work is based on SemEval-2018 dataset for the multi-label emotion classification task “Affect In Twitter” (AIT). Most of this dataset (Arabic section) were collected from social media where dialect is used and since our goal is to classify emotions in MSA language, we built an MSA dataset annotated with emotion MLArEC-1 containing 4,381 sentences, then expanded it with a MLArEC-2 version having 5645 sentences. We applied BERT-based Transfer Learning- method using Arabic pretrained models (BERTbase, ARBERT, MABERT) on our proposed dataset. We conducted an extensive set of experiments and obtained the best results using MABERT with batch size equal to 32 and 10 training epochs and the Micro-F1 rating is 0.94.

Link to full paper

http://journal.damascusuniversity.edu.sy/index.php/engj/article/view/6331/1565